Mac Studio M5 Ultra Makes Local AI Inference Apple's Low-Capex Bet

A desktop built to cut cloud AI bills

Apple's new Mac Studio is the clearest sign yet that the company wants to sell AI compute instead of renting it out. Apple announced the machine on August 25 with two chips, the M5 Max and the new M5 Ultra, and most configurations began shipping on September 22.4 Ars Technica summed up the launch plainly: these desktops were designed around local AI inference and development.2

The spec sheet is mostly about memory. The M5 Ultra is effectively two M5 Max chips joined on one package. It has up to 36 CPU cores, 80 GPU cores, and as much as 512GB of unified memory, and Apple says memory bandwidth reaches about 1.2TB per second.24 Each GPU core now includes a Neural Accelerator, which Apple credits for most of the AI speedup.4 Pricing starts at $2,499 for the M5 Max model and $5,499 for the M5 Ultra.5 One outlet puts the 512GB configuration at roughly $4,000 above the base Ultra price.7

Apple says the hardware is cheaper than paying for cloud AI over time. At launch, the company told enterprise buyers that owning the machines beats renting cloud compute. Hardware chief Johny Srouji called the purchase a one-time investment with no token-based payments.3 To make the point on stage, Apple linked four Mac Studios together, ran a trillion-parameter model to find and fix code vulnerabilities, and powered the whole cluster from one wall outlet.3

Clustering is the real story

The most important change for AI may be the networking, not the chip. Since macOS 26.2 shipped last December, Thunderbolt 5 has supported low-latency links between machines for distributed inference using Apple's open-source MLX framework. Hobbyists and professional researchers responded by chaining Mac minis and Mac Studios together to run models too big for any single consumer machine. Ars described this as a practical alternative to expensive Nvidia GPU servers.2

The new Mac Studio builds that behavior into the product. With RDMA over Thunderbolt 5, several machines can share one pool of memory, and Apple says four units can run inference up to three times faster than one.48 That figure also shows the limit: four machines deliver three times the speed, not four. Clustered Macs are useful for running very large models, but the networking costs some efficiency.

Apple's performance claims vary depending on where you read them. Macworld and other coverage cite up to 4.3 times the peak AI performance of the M3 Ultra.48 Elsewhere, Macworld cites 4.5 times peak GPU compute, and Apple's product page lists a more modest "up to 4x" figure for AI compute.14 Graphics claims differ too, from 1.8 times in one passage to about 40 percent in another.4 One outlet notes that the 4.3x number is Apple's own comparison, not an independent benchmark.8 Buyers should expect the gains to depend heavily on the workload.

Independent testing broadly supports Apple's pitch. PCMag gave the M5 Ultra model an Editors' Choice award and called it a powerhouse for high-end local AI inference, though its review unit cost $12,299 as configured.6 Coverage also disagrees on what the base Ultra includes: PCMag lists 96GB of memory, while another outlet says 256GB comes standard.67 Supply is another catch. Apple promised the 512GB version for late October, but Macworld reports that some orders are slipping to late January.4

How the costs compare with Nvidia

Apple's argument only works when you compare it with what Nvidia hardware costs right now. In October 2026, a standalone Nvidia B200 GPU sells for roughly $45,000 to $50,000 in eight-GPU volumes, up from $30,000 to $40,000 earlier because supply is tight. A B300 runs about $53,000.21 Eight-GPU DGX systems are priced at $280,000 to $350,000, and one reseller lists a DGX B300 at about $647,000.21

Cloud rental prices are also rising. Nebius raised its B300 on-demand price 21 percent on October 1, to $9.50 per GPU-hour.21 B200 rentals range from $5.99 per hour at specialist providers to more than $16 at Google Cloud, and AWS charges $14.24.27 B300 rentals run from $7.10 to $17.80 per GPU-hour.29

Here is a rough comparison, which is our own estimate rather than a figure from Apple. Four maxed-out M5 Ultra Mac Studios at about $9,500 each would cost a little under $40,000. That is less than one standalone B200 and a fraction of a DGX system. The comparison has limits, though. Nvidia's data-center GPUs are built for high-throughput serving to many users at once, and SemiAnalysis benchmarks show the B200 producing tokens at roughly 2.7 to 3.5 times lower cost than the H200 on large models.26 A few Macs on a desk will not replace a production inference cluster. They can, however, replace a lot of the API spending by individual developers, small teams, and companies that need to keep data on site.

Nvidia's next chips will lower cloud prices further. The company says its Rubin platform will cut token costs for agentic, reasoning, and mixture-of-experts workloads to one-tenth of Blackwell's.28 Building AI infrastructure is still expensive: one trade estimate puts a single one-megawatt AI facility at about $60 million.23 Cloud model prices are falling as well. One analysis notes that Apple CEO John Ternus re-promoted the new Macs for AI in the same week as new frontier model launches that were also price cuts.10 Local and cloud inference are both getting cheaper, and Apple is betting that a one-time purchase will appeal to buyers more than per-token billing does.

Why investors care: customers pay for the compute

The Mac Studio fits neatly into Apple's broader strategy of spending very little on AI infrastructure, which investors have rewarded. One report puts Apple's fiscal 2026 capital spending at $14 billion. That compares with about $650 billion that Alphabet, Amazon, Meta, and Microsoft plan to spend on AI data centers this year.11 Estimates of the gap vary. The Motley Fool counts $6.8 billion in Apple capex over three fiscal quarters, against roughly $800 billion targeted by five hyperscalers, a figure that comes from Nvidia CFO Colette Kress.18 Another analysis compares Apple's estimated sub-$15 billion with Alphabet's guidance of $195 billion to $205 billion.15 Every count shows a gap of more than ten times.

Apple also avoided building its own frontier model. It reportedly pays Google about $1 billion a year for a custom Gemini model of roughly 1.2 trillion parameters, which powers the rebuilt Siri.15 Meanwhile, Apple bought back $90.7 billion of its own stock in fiscal 2026.11 Its June-quarter revenue rose 16 percent to $109.4 billion.13

The Mac Studio takes that approach one step further. One analysis puts it simply: hyperscalers buy inference capacity and depreciate it, while Apple sells inference capacity and records it as revenue at about a 50 percent gross margin.15 Each Mac Studio sold adds AI compute to Apple's ecosystem, and the customer pays for it. Mac revenue rose 29 percent in the June quarter to $10.4 billion, a record for that quarter, which suggests the strategy is helping sales and not just shaping the story.15 The stock reflects this. Apple is up roughly 25 percent this year, the best performer among the Magnificent Seven and ahead of Nvidia.1516

The risks

The bullish case leaves out some real weaknesses. Apple depends on Google, a direct competitor, for a core Siri feature. The $1 billion fee may be a promotional rate that won't last at renewal, and regulators in the US and EU could force Apple to offer users a choice of assistant.1516 On the hardware side, Citi warns that rising memory prices are squeezing gross margins.20 That matters for a product whose main selling point is a huge pool of memory. If memory stays expensive, the 512GB configurations that make the clustering pitch work will either get pricier or earn Apple less.

There is also a larger question. If AI infrastructure turns out to be the decisive investment of the decade, Apple's spending discipline will look like a strategic mistake that flattered its cash flow.19 That risk is real, but the current evidence favors Apple. Graphics processor prices and cloud rental rates are both rising, so the Mac Studio's one-time price looks better each month for developers running large open models. Apple is not trying to beat Nvidia in the data center. It is trying to win the AI workloads that never need to leave a customer's desk, and for now, buying hardware for those jobs looks cheaper than renting it.